Manufacturing AI must move past experimentation
Manufacturing leaders are no longer short of AI ideas. Every function can produce a list of possible uses, from productivity tools and maintenance copilots to quality analysis, process optimisation and administrative automation. The harder question is whether the organisation can decide which problems deserve AI, which should be handled through conventional analytics or process improvement, and which experiments should be stopped before they consume attention that should be focused elsewhere.
Manou Campbell, head of data and information systems at Medik8, says manufacturers have entered a more demanding phase. The early period of broad AI exploration helped companies understand what the tools could do, but it did not automatically create a strategy.
“I think we are past the age of experimentation,” he says. “The last few years were about working out the generic use of AI, how it can be helpful, how we can use it and where we can see the value. Now, although it is useful to look for situations where we can help productivity, it is within the processes that we need a more specific solution that is really starting to show value. At Medik8, our approach is to find specific use cases, apply the resources there, and then we get a much better uptake or engagement from those elements.”
That discipline matters because manufacturing AI can touch production, safety, quality, supply chain performance and workforce productivity. A use case that looks attractive in isolation still must compete for resource, governance and operational attention against other priorities.
AI is not always the answer
Campbell argues that manufacturers should be careful not to treat generative AI as the default response to every performance problem. Many issues can still be solved through conventional analytics, automation, machine learning or a change to the underlying process. The first decision is not which AI tool to use, but whether the problem requires AI at all.
“There are nuances between analytics, machine learning and AI, and manufacturers had a lot available for process optimisation before the use of generative AI,” he says. “The question is where the generative aspect adds value compared to what we already had. In manufacturing, the main thing we are typically trying to do is save cost or improve productivity, and most of those things can be improved through traditional methods such as machine learning or understanding where there are inefficiencies. Where generative AI is a better solution is in understanding areas that we may not have considered, or where we do not know where to start.”
A known bottleneck on a production line may need analysis of the constraint, not a generative AI deployment. A broader question about where performance is being lost, or what improvement area has not yet been considered, may be better suited to tools that can interrogate information more flexibly.
Dashboards and single operational views still have value when the business already knows what it wants to monitor, but they are less useful when the question changes or when a leader needs to understand the reasons behind a performance shift. Campbell argues that AI becomes more valuable when it helps people move beyond the predefined report, reach the underlying information faster and ask better follow-up questions without waiting weeks for a data team to return an answer.
“With the traditional method, you have one view where you can access all your data in one neat place,” Campbell says. “It is a predefined report or collection of data, and you see it. But the data underneath that answer can now be accessed a lot more quickly with AI. Traditionally, if you had that question and went to a data team, you might have to wait a couple of weeks to get the information back. Now you can ask a prompt that has access to it and get the information within a few minutes. You want more information underneath it to give you the why, rather than just what is happening.”
Start with the objective
The weakest AI projects often begin with enthusiasm rather than intent. Before approving an initiative, Campbell says leaders should be clear about the objective, success criteria, test method, control group and alternatives being compared. “For me, before you even have a question, what is your objective?” he says. “If you are leading a project like this without an objective, you are already starting off on the wrong foot. Then the next questions should be: what is your success criteria, how are you testing it, who is using it and why are they using it? What is your control group versus your alternatives? With AI, are you testing it against a traditional method first, is it something that stands in its own right, or are there different versions of AI to test?”
Those questions make governance more practical because they connect oversight to the purpose of the project. They also make it harder to confuse activity with progress. A pilot can produce an impressive demonstration and still fail if it does not improve cost, output, safety, admin time, labour productivity or the ability to grow without equivalent headcount.
Once an AI initiative is in use, someone also has to own it, maintain it, check it, manage security and compliance, and decide how far it should be allowed to act. Poor data quality is often treated as a reason to delay AI, but Campbell says the right threshold depends on the use case. “This comes down to your appetite as an organisation to run with a minimum viable product, or something that is 80 percent of the way there, versus waiting for perfection,” he says. “If you can reliably trust the data your company is producing, and it has most of the sources in there, then you can release the product you have and spend the rest of your time iteratively building on top. The alternative is that you get as close to perfection as possible, plugging in all your sources and making sure there is stringent compliance across your data layers. It depends on the significance of that data in manufacturing.”
Scale depends on discipline
Manufacturers may want rapid progress, but Campbell argues that reliable integration, ingestion, transformation, orchestration and testing must come before AI is asked to support output or decisions. “If people are trying to skip over the traditional or modern sense of building a reliable data stack, then they are missing the core foundation,” he says. “You should have a very modernised, efficient system underneath for how you integrate your data, how you ingest it, how you transform it into a reliable sense and how you orchestrate it through the process. Only when you have an incredibly reliable stack underneath should you plug in AI on top to refer output or make decisions based on that data.”
AI can also make data teams faster. Campbell says models that once took weeks can now be built in hours or days, with AI helping on early validation and testing. The constraint then becomes prioritisation. “At the moment, everyone in data and tech can feel like a kid in a sweet shop because there are so many options and so many exciting projects to undertake,” he says. “It is about being incredibly strict, from top to bottom. What are your company objectives? What are the tech and data objectives that feed into those company objectives? From that, what can be solved by AI, or what do we need to experiment with AI to solve? Does that hold true all the way through, so we are still moving toward the objectives of the company and not getting sidetracked by things that are now possible?”
A proof of concept should scale only when it has met the original objective, shown user engagement, passed security and compliance requirements, and proved it can be maintained without excessive overhead. It should not continue simply because time has already been invested. “I think areas are often born where there is not a clear controlled pilot phase, or where there are too many pilot initiatives tested out on a wide scale without a very clear objective or success criteria,” Campbell says. “There needs to be a clear framework within the organisation of what is being tested, why and what the output is. You do not want to get caught in the idea that just because you built it, there is value in it. Even if you have invested time, if you are not seeing the output, then it is time to pivot or remove the initiative.”
Campbell will explore these issues further at the Manufacturing Data Summit UK, which takes place in London on 6 October 2026 and brings together manufacturing leaders to discuss AI, analytics, data quality, governance and operating models. For Campbell, the value lies in hearing how peers are approaching the same problems. “We do not know what we do not know,” he says. “With shared ideas, networking and hearing panels, there are always things I can go away with and explore, and that drives conversations back in my own team.” Register now to join the discussion and secure your place at the summit.
